Computer Agents Your Team Can Share, Channel Backgrounds & AI Priced by the Week

Computer agents now live on a machine of their own: set one up, invite teammates into its channel to talk to it, and keep the computer itself private to you. Every channel can be drawn as a list or as bubbles over a preset, a colour, a gradient or your own picture. AI stopped being priced in dollars and is now a share of your week, with a new Billing home, a message that is saved and sent for you when your week reopens, and the option to pause your plan instead of cancelling. And a device sign-in can no longer be approved by simply opening a link.

New Features

Computer Agents — An Agent With Its Own Computer

A computer agent is an agent that works on a real machine: it opens the browser, writes files, runs what it needs and reports back in your channel.

• Set one up in a few steps — name it, choose its computer, decide how much AI it may use and describe what it should do. The runtime installs itself on the machine and verifies it is ready before the agent starts working.
• A computer panel beside the conversation — watch the machine while the agent works, with chat and computer as tabs in the channel header. On narrow screens the computer opens as an overlay over the chat.
• Your computers as cards, not a dropdown — each card names who can open the machine and whether it can be used. A computer that is busy, reserved or already bound stays visible with the reason it cannot be chosen.
• Choosing a shared computer asks first — picking a machine other people can open tells you plainly that setting it up makes it yours alone, names exactly who loses access, and asks again if that list changed while you were deciding. Everyone removed is told, and their open sessions are closed.
• Approvals for anything that matters — the agent asks before actions it should not take on its own, and its approval cards cannot be reused for a different machine.
• Installed in your language — a new machine takes the language, timezone and keyboard layout of the person who bought it or created the agent, including the browser's own language, and keeps them when ownership changes.
• It works from your email too — an agent that sent an email wakes up when someone replies to it, including threads that get split, and never reacts to your customers' ordinary incoming mail.
• Voice, pictures and formatting — voice messages are transcribed for the agent, pictures and videos appear as framed media in the conversation, and the agent's replies arrive as properly formatted text instead of raw markup.

Computer Agents — Teammates Can Join, and the Computer Stays Yours

Inviting someone into a computer agent's channel now shares the agent, exactly as it does for a Copera agent.

• "Who has access to this agent" — pick the teammates who may talk to it when you create it, and add, remove or change roles later from the channel's member settings. A plain warning says what everyone added can do.
• The agent acts with your access — a teammate's message wakes it, but it still works with the permissions of the person who created it.
• The machine is not shared — teammates get no computer tab, no computer settings, no reinstall and no console. They can see that it is setting up or offline, and Stop and Deny work for everyone.
• The creator cannot be removed — not by a channel admin, not by workspace-wide agent administration, and the creator cannot leave their own agent's channel.
• Roles decide who gets a computer — a role now carries whether its holders may use computers and which controls they get, resolved once for everyone in a channel instead of per message.
• Agents stay where you put them — a computer agent dragged into a sidebar category stays in it, across reloads and on every open client.

Computer Agents — Stop Them Replying on WhatsApp

A computer agent that had answered once in a WhatsApp conversation woke up on every message the contact sent, and nothing the owner did could stop it.

• Pause and resume, per conversation — a control in the conversation header pauses the agents in that thread and a banner says so, replacing the old and inaccurate "bot" banner. Pause and resume both show up in the conversation's activity timeline, with who did it.
• Or switch a whole channel off — a channel setting decides whether agents may reply on it at all, independently of your native bot, so resuming agents never switches a bot back on.
• Replying yourself pauses them — a message you send from Copera pauses the agents in that thread, so two answers never race to the customer.
• Replying from your phone also takes over — a reply typed in WhatsApp itself now stands your bot down, pauses the agents and parks the conversation as waiting on the customer, with the first response time recorded. Copera's own outgoing messages are recognised as echoes and never mistaken for a takeover, and a reconnect that backfills old history never pauses an inbox.
• The agent is told why it is quiet — each time a conversation or a channel is paused, the agent gets one note in its own channel explaining that a contact's message did not reach it, in its own language.

Chat — Give Every Channel Its Own Look

Every channel can now be shown as a list or as bubbles, over a background you choose (#3908).

• Six kinds of background — none, the Classic, Aurora or Paper presets, a solid colour from a picker or your theme's swatches, a two-colour gradient with an angle, or your own picture (JPEG, PNG or WebP up to 5 MB, uploaded as a private file).
• For everyone, or just for you — people who can manage the channel set its default; anyone can set a personal look that wins over it, with Reset to channel default to go back. A note tells you when you are using your own.
• A live preview — the new Appearance page in channel settings previews your choice as you make it, and every change applies at once.
• Readable on anything — text, day separators, notices, reply chips and avatars keep a readable contrast on every background and in every theme, and the bubble view shows your background at full strength instead of dimming it behind a pale lane.
• Media framed like a messenger — pictures, GIFs and videos sit in a coloured frame inside the bubble, with the time as a pill over the media and the caption underneath. Voice messages get a bubble of their own, and the web player now matches the mobile one.
• Available everywhere — text channels, direct messages, agent and computer agent channels, meeting rooms and classrooms. Threads follow their parent channel's look.

Compute Consoles — Shared Terminal Tabs

• More than one terminal per console — open several terminal tabs on the same machine, name them, rename and close them, with each tab getting its own session.
• Shared with the other people in the console — a tab someone opens appears for everyone, and closing it on your side keeps it open on theirs.
• Links open in your browser — Ctrl or Cmd + click a web address printed in the terminal and it opens in your own browser, not the machine's.
• Copy and paste that behaves — Ctrl/Cmd+Shift+C and V copy and paste locally while plain Ctrl+C still interrupts the remote program, and a selection made on the machine syncs back to you.
• Compute channels in your sidebar categories — a machine's channel can be filed like any other, and opening its settings no longer reconnects the console.

AI Plans — Priced by the Week, Not in Dollars

How much AI you have is now expressed as a share of your week everywhere it appears, and the whole plan, usage and payment experience was rebuilt around it (#3862).

• Share of the week, not cents — budgets, the AI usage page, automation usage, agent runs and approvals, hiring from the store, meeting actions and Ask Copera all read as a percentage of the week. Payments still read in money.
• A Billing home — Overview, Usage, Add-ons, Budgets and Payments as tabs. The overview shows your plan, the last four weeks as a chart against the 100% line, extra usage, your add-ons, the next invoice and your payment summary. The old credits, automation usage and AI usage pages redirect into it.
• Your message is saved, not lost — a message refused because the week is used up is held and sent automatically the moment AI comes back: an upgrade, extra usage, a refill, or simply the weekly reset. You can ask to be notified when it runs.
• Ask an admin for more — members can request more AI with a note; admins see the requests, answer them, and an upgrade, top-up or refill answers them automatically. Requesters are told who answered.
• Quiet until it matters — nothing is shown below 80% of your week. From 80% you get a forecast, a line at 95%, and a calm banner you can dismiss. Admins can opt into an 80% email and get a used-up email by default, plus an optional weekly summary.
• Upgrade without leaving your work — every upgrade now opens as a dialog where you already are (the limit card, the banner, the usage sheet, the billing overview) with the picks, the exact amount due today, the card check and the bank's confirmation, and the week refreshes in place. The dialog cannot close while a charge is running.
• Picks that span the ladder — the sheet offers a cheaper option, the recommendation (preselected, and chosen from how much AI your last four weeks actually needed) and the largest plan, with how often the cheaper one would have run out. Every bar of "the bigger the plan, the cheaper the AI" selects its plan.
• Change plan, with history — the change-plan page replays your last four weeks against the plan you are looking at, with price and discount per row and a scheduled downgrade you can undo.
• Pause instead of cancelling — pause your plan for one or two months from the end of the paid period. You keep working on the free tier and the paid plan returns on the next paid renewal. Resume before it starts costs nothing, Resume now charges today. Reminder emails arrive three days before a plan ends and seven days before a pause ends.
• A cancel flow that listens — pick a reason and you may be offered a right-sized plan, a pause, extra usage with a cap, or a discount, then a facts-only confirmation with nothing preselected and an undo at the end.
• Payment methods and recovery — add a card, make it the default, or remove it with a replacement picker (a card still paying a live subscription cannot be removed). A failed payment can be retried with a backup card or confirmed with your bank on the original one.
• Extra usage and auto top-up — packs are labelled by the share of a week they buy, with an upgrade nudge only when a bigger plan is actually cheaper than topping up. Auto top-up was being charged without crediting the workspace — that is fixed, with declines backing off and admins notified before it is switched off.
• Add-ons — a live prorated preview, and the amount you are quoted is the amount you are charged. Increasing an add-on prorates; decreasing or cancelling one is booked for its own renewal, so you keep what you paid for, with "ends on" and Undo shown until the date.
• Free to paid — the free limit card offers two plans and nothing else, and a free workspace that ran out of AI in two of its last four weeks gets 50% off its first month for seven days. Referrals earn both sides a banked week refill when the invited workspace pays its first plan invoice.
• In the inbox — thread summaries and reply suggestions no longer just fail when AI is paused: you get a calm "paused until your week resets" bar with the actions your role allows, and "AI is busy" while work already running finishes.
• Saved cards and honest dates — a saved card that needs your bank's confirmation now opens the check on that card instead of showing a blank card form, a declined payment says it was declined with "Use another card", and a first purchase quotes your renewal one month from today rather than a stale date (#3891).

My Work — Tick Rows Off for Yourself

• Personal completion — complete or reopen a row in My Work without touching the source row or anyone else's view. Your completion applies to lists, filters and counts, with instant feedback and a rollback if the save fails.

Boards — Find a Row by Its ID or Its Words

• Search across readable columns — the grid searches text, paragraph and description cells, as well as row IDs, with your view's filters and the search scope kept across pages and exports.

Desktop — Keep the Screen Awake

• A new device setting stops your display going to sleep while Copera is open, restored the next time you start the app.

Bug Fixes

• Device sign-in approved by merely opening a link — opening a desktop sign-in link while signed in used to approve it on page load. The page now only shows the request (where it came from, when, and the device it reports) with a countdown and Approve or Cancel; the same account must approve, a different network also requires the 4-digit code shown on the requesting screen, requests expire after five minutes, and the new device's login history records who approved it (#3903).
• Sign-in tokens issued for another app — a Google or Apple token minted for a different site could be used to sign into a Copera account. Tokens are now checked against Copera's own apps. Personal access tokens also stop working when a member leaves, is removed or disabled, and when the account password is reset, changed or set (#3901).
• "Just signed in" after an ordinary action — creating a workspace, accepting an invite, changing a password or approving a sign-in no longer makes a session look freshly authenticated, so checks that require a recent sign-in cannot be passed by an old session (#3904).
• Customer content in product analytics — session recordings are off, analytics no longer collect the text or attributes of what you clicked, and page titles (which carry channel, file, board and contact names) are no longer sent (#3899).
• Advertising tags loading with no campaign running — ad and analytics tags on copera.ai and in the app now load only while the platform is switched on and your cookie choice allows it, with the cookie banner and settings shown only while something optional is actually running. Global Privacy Control is honoured (#3900, #3907).
• Transcription speakers and deletion — naming a speaker and deleting a transcript are separate rights now, and the speaker picker keeps its roster instead of emptying (#3924).
• Call duration and history dates on direct calls — the duration and the date shown in your call history are correct, including when someone leaves early or a duration is missing (#3922).
• Camera indicators in meetings no longer disagree about who has their camera on (#3898).
• Guests in meetings — admitted guests connect reliably again, their comments load with the right state and contrast, and a guest session carries through the voice context. Incomplete admissions are refused instead of half-joining (#3882, #3883).
• Guests who already left showing in external-user warnings — the warning now counts only guests who are still active.
• Chat messages that felt slow to send — text is sent immediately, your draft is restored if the send is refused or fails, repeated Enter presses and Enter while composing with an input method are ignored, and a message is no longer sent twice on a retry.
• Voice replies on WhatsApp losing their context — an audio reply keeps the message it was answering (#3919).
• GitHub handoffs in Code Review — connecting opens in your real browser, retries keep working, and the connection can only be completed by the browser that started it. Automatic reviews and approvals now verify that the contributor actually has write access first, and an approval is no longer dismissed when the pull request has not changed (#3923, #3880).
• Agents reporting work they had not finished — results are attached natively, outcome feedback is explicit, tool work is properly bounded, transcript identity is correct and permissions are checked on the fallback path (#3886).
• "Missing file" on a PDF you did supply — a document request is checked against the response actually asked for, admitted PDF sources are kept whole, and completeness is only judged at the end (#3920).
• Transcript links and channels — transcripts link to the right place, non-member administrators see their sources refreshed, and transcription access is consistent (#3925).
• Tempo on the day the clocks go back — an hour that happens twice now always resolves to the earlier one, so the day's axis no longer depends on when you look at it (#3934).
• Rows pointing at deleted members or linked rows no longer break the board (#3888).
• Meeting, Drive and task regressions — including expired invites being offered in a connection handoff, and invite status that disappeared instead of reporting the problem (#3889).
• Leaving a Design preview closed the desktop workspace behind it (#3895).
• .docx files failing to convert — importing a Word file into Docs and converting one for signature work again (#3890).
• Drive search with a star or other special characters returned an error instead of results (#3890).
• Images that failed to resize now fall back to the original instead of showing nothing (#3894).
• Dialogs crashing after a new version shipped — a dialog whose code went stale reloads itself once instead of showing an error (#3894).
• Lookups, rollups and sorting — lookups chained through other lookups no longer report false drift, sorting by several multi-value columns at once works, and empty cells are skipped during reconciliation (#3896).
• Editor crashes — pasting into a composer with no upload option, highlighting a selection, emoji and mention fallbacks, and opening a brand-new thread all stopped crashing (#3894, #3896, #3879).
• Microphone and camera permissions — a denied prompt or a machine with no microphone now shows a notice instead of a crash (#3879, #3896).
• Captions for words the provider timed at a single instant are now generated instead of dropped (#3892).
• Notifications and copy — new-email notifications name the inbox, import reports show their limits, documents without a title read as "Document" instead of blank, and the delete dialog names the integration you are removing (#3894).
• WhatsApp messages to a number with no account are reported as undeliverable instead of retried forever (#3890).
• Push notifications to a stale device are dropped rather than retried (#3902).
• Inbox and mail server outages — folders, threads and attachment downloads now fall back or retry quietly when the mail server is briefly unavailable, instead of surfacing errors, and inboxes created before a change in how aliases are stored list correctly again (#3890, #3892, #3894).
• Public chat history — asking for messages before or after a date, instead of a message, answers with a clear error rather than failing (#3892).
• Automation and document templates — specific save errors are preserved instead of being replaced by a generic one, and template permissions are checked in batches so large libraries load (#3874).

Improvements

• Computers in Workspace Settings — compute managers get a workspace-wide inventory of machines, can administer who belongs to each one and transfer ownership, while an agent's private computer stays private. Removing a member revokes their console access, and setup only asks for attention when something is genuinely blocking it.
• Privacy policy and security contact — the policy now states that data is stored in the United States (with rented compute machines as the one exception), explains that deleting a workspace also deletes its signed documents, describes what an e-signature audit trail records, and publishes a dedicated address for security reports (#3910).
• A new copera.ai — the marketing site was rebuilt around one argument: the workspace is free and intelligence is what you pay for, at one flat price. Every page follows the new design, with cookie consent that opts you in or out according to where you are, a shareable seat-tax calculator, a "why it's free" essay and a download page that offers the right mobile app for your device. All six languages (#3897, #3905).
• Legal pages say only what Copera does today — claims about certifications, deletion, export, encryption and healthcare were corrected or removed across all six languages, and e-signature wording no longer claims universal legal validity (#3897).
• AI that fails over on its own — Copera now reaches several AI providers directly rather than always through one route, so a provider outage, a dead key or exhausted credits moves your request to the next option mid-answer instead of ending it. Sticky sessions return to their original provider once it recovers (#3909, #3918, #3926).
• Reasoning models that think for longer — a model that spends a long time reasoning is no longer cut off as idle, with a firm ceiling so a genuinely stalled answer still ends (#3926).
• Ask Copera answers in prose when it has to — if a reply arrives without its expected structure, the text is recovered and shown with its tool steps intact, instead of failing (#3894).
• Fewer phantom errors and faster recovery — a large pass over the whole product made expected situations (a brief network drop, a job that will retry, a shutdown mid-task) stop surfacing as failures, so the errors you do see are real ones (#3878, #3884, #3885, #3890, #3892, #3894, #3896, #3902).
• Calendar invitations with unusual titles no longer break the agenda, and an organiser address that refuses mail is reported quietly (#3892, #3894).
• Email sent to addresses that cannot receive it is skipped before it leaves, and signing-request failures are no longer reported twice (#3892).